Using magnetic resonance imaging to distinguish a healthy brain from a bipolar brain: A transfer learning approach

Bipolar Disorder (BD) is a recurrent psychiatric condition characterised by periods of depression and (hypo)mania, it affects more than 1% of the world's population [1]. However, accurate diagnosis can be difficult due to the lack of diagnostic tools available to practitioners. To address this knowledge gap this paper aims to understand how the application of transfer learning, in the context of machine learning techniques, can be used to improve a diagnosis of BD. Image detection of magnetic resonance images (MRI) was undertaken to identify features of grey matter in BD brains in comparison to healthy controls (HC), which may constitute a biomarker of BD. Additionally, the products of machine learning were investigated for clinical application to efficiently aid in clinical diagnosis by an end user, through a cloud-based application. The transfer learning model created demonstrated at 88% accuracy the ability to detect features present in the BD brain, not present in controls. Of limitation to this study was the amount of MR images required to train this model. However, this project identifies that it is possible with limited resources to create a model which may prove useful in diagnostic settings in the future.